The creator of the Agent Auth Protocol argues that AI assistants need their own identities, scoped authority, audit trails, and a kill switch before unsafe security habits become permanent.